- tidying up code

- simplifying graph-context expecting both dense weights

Author: sfallah

src/llama-graph.cpp

@@ -1856,18 +1856,15 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
 void llm_graph_context::build_dense_out(
     ggml_tensor * dense_2,
     ggml_tensor * dense_3) const {
-    if (!cparams.embeddings) {
+    if (!cparams.embeddings || dense_2 == nullptr || dense_3 == nullptr) {
         return;
     }
-    ggml_tensor * cur = res->get_embd_pooled();
-    if (dense_2 != nullptr) {
-        cur = ggml_mul_mat(ctx0, dense_2, cur);
-        cb(cur, "result_embd_pooled", -1);
-    }
-    if (dense_3 != nullptr) {
-        cur = ggml_mul_mat(ctx0, dense_3, cur);
-        cb(cur, "result_embd_pooled", -1);
-    }
+    ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd;
+    GGML_ASSERT(cur != nullptr && "missing t_embd_pooled/t_embd");
+
+    cur = ggml_mul_mat(ctx0, dense_2, cur);
+    cur = ggml_mul_mat(ctx0, dense_3, cur);
+    cb(cur, "result_embd_pooled", -1);
     res->t_embd_pooled = cur;
     ggml_build_forward_expand(gf, cur);
 }

src/llama-model.h

@@ -440,8 +440,8 @@ struct llama_model {
     //Dense linear projections for SentenceTransformers models like embeddinggemma
     // For Sentence Transformers models structure see
     // https://sbert.net/docs/sentence_transformer/usage/custom_models.html#structure-of-sentence-transformer-models
-    ggml_tensor * dense_2_out_layers;
-    ggml_tensor * dense_3_out_layers;
+    struct ggml_tensor * dense_2_out_layers = nullptr;
+    struct ggml_tensor * dense_3_out_layers = nullptr;
 
     llama_model_params params;